The records behind this event
2-
Chip
NVIDIA B200
Blackwell: two reticle-limit dies fused into one GPU, the 2025 workhorse.
Blackwell hit the reticle limit and kept going: the B200 joins two 104-billion-transistor dies with a 10 TB/s die-to-die link so they behave as one 208-billion-transistor GPU. It introduced FP4 precision for inference, moved to CoWoS-L packaging, and in its rack-scale GB200 form made liquid cooling mandatory at the frontier. Every major cloud deployed it…
Open record → -
System
NVIDIA GB200 NVL72
The rack that is the computer: 72 GPUs, one NVLink domain, 120 kW, liquid only.
The NVL72 abandoned the server as the unit of design: 18 compute trays (36 Grace CPUs + 72 B200 GPUs) and 9 NVLink-switch trays joined by a copper spine into a single 72-GPU coherent domain with 13.4 TB of HBM3E, presented to software as one enormous accelerator. At ~120 kW it made direct liquid cooling and 1,400 A busbars mandatory, forcing the data-center…
Open record →